Skip to content
← Models
Open weightsBlack Forest Labs/10 trustVerified
FLUX.1 Kontext logo

FLUX.1 Kontext

Instruction-based image editing that keeps character and style consistent across edits.

Visit official page →

About

FLUX.1 Kontext, announced May 29, 2025, extends the FLUX line from pure generation into in-context editing: give it a text instruction plus a reference image and it applies targeted edits — object changes, lighting shifts, partial scene transforms — while keeping the rest of the image and the subject's identity intact. The suite ships as Kontext [pro] for iterative editing workflows and Kontext [max] for maximum prompt adherence and typography, both API-only, plus an open-weight 12B Kontext [dev] under a non-commercial license. BFL claims inference up to 8x faster than leading editing models at release and reports top scores on its internal KontextBench for text editing and character preservation. Adobe's Photoshop beta adopted Kontext Pro for generative fill, which is about as strong a production endorsement as an editing model gets. Limitation to note: multi-turn edits can still drift on fine identity details, and the dev weights forbid commercial use.

Specs

Context window

Pricing

Kontext [pro] ~$0.04/image, Kontext [max] ~$0.08/image via the BFL API

Free quota

API access

Yes — BFL API, Replicate, fal.ai, Together AI, Krea, Freepik, Leonardo

Weights

Open weights

Released

May 29, 2025

Benchmarks

BFL-reported: top scores on internal KontextBench for text editing and character preservation (provider-reported)

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

image editingin-context generationcharacter consistencylocal edits

Trust score — 7/10

/10 trustVerified

Scored 7/10 (verified tier, metrics-v1). Evidence: Announced May 29, 2025 per BFL/Wikipedia; suite: Kontext [pro], [max] (API) + [dev] open weights 12B (non-commercial).; Adobe Photoshop beta adopted Kontext Pro for generative fill — strong production signal.; BFL claims up to 8x faster inference vs leading editing models; top scores on internal KontextBench (provider-reported).; Limitations: multi-turn identity drift possible; dev weights non-commercial only..

Sources